Throughput Benchmark Sample

This sample demonstrates how to estimate performance of a model using Asynchronous Inference Request API in throughput mode. Unlike demos this sample does not have other configurable command-line arguments. Feel free to modify sample’s source code to try out different options.

The reported results may deviate from what benchmark_app reports. One example is model input precision for computer vision tasks. benchmark_app sets uint8, while the sample uses default model precision which is usually float32.

Before using the sample, refer to the following requirements:

How It Works

The sample compiles a model for a given device, randomly generates input data, performs asynchronous inference multiple times for a given number of seconds. Then, it processes and reports performance results.

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
# Copyright (C) 2022 Intel Corporation
# SPDX-License-Identifier: Apache-2.0

import logging as log
import sys
import statistics
from time import perf_counter

import numpy as np
import openvino as ov
from openvino.runtime import get_version
from openvino.runtime.utils.types import get_dtype


def fill_tensor_random(tensor):
    dtype = get_dtype(tensor.element_type)
    rand_min, rand_max = (0, 1) if dtype == bool else (np.iinfo(np.uint8).min, np.iinfo(np.uint8).max)
    # np.random.uniform excludes high: add 1 to have it generated
    if np.dtype(dtype).kind in ['i', 'u', 'b']:
        rand_max += 1
    rs = np.random.RandomState(np.random.MT19937(np.random.SeedSequence(0)))
    if 0 == tensor.get_size():
        raise RuntimeError("Models with dynamic shapes aren't supported. Input tensors must have specific shapes before inference")
    tensor.data[:] = rs.uniform(rand_min, rand_max, list(tensor.shape)).astype(dtype)


def main():
    log.basicConfig(format='[ %(levelname)s ] %(message)s', level=log.INFO, stream=sys.stdout)
    log.info('OpenVINO:')
    log.info(f"{'Build ':.<39} {get_version()}")
    device_name = 'CPU'
    if len(sys.argv) == 3:
        device_name = sys.argv[2]
    elif len(sys.argv) != 2:
        log.info(f'Usage: {sys.argv[0]} <path_to_model> <device_name>(default: CPU)')
        return 1
    # Optimize for throughput. Best throughput can be reached by
    # running multiple openvino.runtime.InferRequest instances asyncronously
    tput = {'PERFORMANCE_HINT': 'THROUGHPUT'}

    # Create Core and use it to compile a model.
    # Select the device by providing the name as the second parameter to CLI.
    # It is possible to set CUMULATIVE_THROUGHPUT as PERFORMANCE_HINT for AUTO device
    core = ov.Core()
    compiled_model = core.compile_model(sys.argv[1], device_name, tput)
    # AsyncInferQueue creates optimal number of InferRequest instances
    ireqs = ov.AsyncInferQueue(compiled_model)
    # Fill input data for ireqs
    for ireq in ireqs:
        for model_input in compiled_model.inputs:
            fill_tensor_random(ireq.get_tensor(model_input))
    # Warm up
    for _ in range(len(ireqs)):
        ireqs.start_async()
    ireqs.wait_all()
    # Benchmark for seconds_to_run seconds and at least niter iterations
    seconds_to_run = 10
    niter = 10
    latencies = []
    in_fly = set()
    start = perf_counter()
    time_point_to_finish = start + seconds_to_run
    while perf_counter() < time_point_to_finish or len(latencies) + len(in_fly) < niter:
        idle_id = ireqs.get_idle_request_id()
        if idle_id in in_fly:
            latencies.append(ireqs[idle_id].latency)
        else:
            in_fly.add(idle_id)
        ireqs.start_async()
    ireqs.wait_all()
    duration = perf_counter() - start
    for infer_request_id in in_fly:
        latencies.append(ireqs[infer_request_id].latency)
    # Report results
    fps = len(latencies) / duration
    log.info(f'Count:          {len(latencies)} iterations')
    log.info(f'Duration:       {duration * 1e3:.2f} ms')
    log.info('Latency:')
    log.info(f'    Median:     {statistics.median(latencies):.2f} ms')
    log.info(f'    Average:    {sum(latencies) / len(latencies):.2f} ms')
    log.info(f'    Min:        {min(latencies):.2f} ms')
    log.info(f'    Max:        {max(latencies):.2f} ms')
    log.info(f'Throughput: {fps:.2f} FPS')


if __name__ == '__main__':
    main()

// Copyright (C) 2022 Intel Corporation
// SPDX-License-Identifier: Apache-2.0
//

#include <algorithm>
#include <condition_variable>
#include <string>
#include <vector>

// clang-format off
#include "openvino/openvino.hpp"

#include "samples/args_helper.hpp"
#include "samples/common.hpp"
#include "samples/latency_metrics.hpp"
#include "samples/slog.hpp"
// clang-format on

using Ms = std::chrono::duration<double, std::ratio<1, 1000>>;

int main(int argc, char* argv[]) {
    try {
        slog::info << "OpenVINO:" << slog::endl;
        slog::info << ov::get_openvino_version();

        std::string device_name = "CPU";
        if (argc == 3) {
            device_name = argv[2];
        } else if (argc != 2) {
            slog::info << "Usage : " << argv[0] << " <path_to_model> <device_name>(default: CPU)" << slog::endl;
            return EXIT_FAILURE;
        }
        // Optimize for throughput. Best throughput can be reached by
        // running multiple ov::InferRequest instances asyncronously
        ov::AnyMap tput{{ov::hint::performance_mode.name(), ov::hint::PerformanceMode::THROUGHPUT}};

        // Create ov::Core and use it to compile a model.
        // Select the device by providing the name as the second parameter to CLI.
        // It is possible to set CUMULATIVE_THROUGHPUT as ov::hint::PerformanceMode for AUTO device
        ov::Core core;
        ov::CompiledModel compiled_model = core.compile_model(argv[1], device_name, tput);
        // Create optimal number of ov::InferRequest instances
        uint32_t nireq = compiled_model.get_property(ov::optimal_number_of_infer_requests);
        std::vector<ov::InferRequest> ireqs(nireq);
        std::generate(ireqs.begin(), ireqs.end(), [&] {
            return compiled_model.create_infer_request();
        });
        // Fill input data for ireqs
        for (ov::InferRequest& ireq : ireqs) {
            for (const ov::Output<const ov::Node>& model_input : compiled_model.inputs()) {
                fill_tensor_random(ireq.get_tensor(model_input));
            }
        }
        // Warm up
        for (ov::InferRequest& ireq : ireqs) {
            ireq.start_async();
        }
        for (ov::InferRequest& ireq : ireqs) {
            ireq.wait();
        }
        // Benchmark for seconds_to_run seconds and at least niter iterations
        std::chrono::seconds seconds_to_run{10};
        size_t niter = 10;
        std::vector<double> latencies;
        std::mutex mutex;
        std::condition_variable cv;
        std::exception_ptr callback_exception;
        struct TimedIreq {
            ov::InferRequest& ireq;  // ref
            std::chrono::steady_clock::time_point start;
            bool has_start_time;
        };
        std::deque<TimedIreq> finished_ireqs;
        for (ov::InferRequest& ireq : ireqs) {
            finished_ireqs.push_back({ireq, std::chrono::steady_clock::time_point{}, false});
        }
        auto start = std::chrono::steady_clock::now();
        auto time_point_to_finish = start + seconds_to_run;
        // Once there’s a finished ireq wake up main thread.
        // Compute and save latency for that ireq and prepare for next inference by setting up callback.
        // Callback pushes that ireq again to finished ireqs when infrence is completed.
        // Start asynchronous infer with updated callback
        for (;;) {
            std::unique_lock<std::mutex> lock(mutex);
            while (!callback_exception && finished_ireqs.empty()) {
                cv.wait(lock);
            }
            if (callback_exception) {
                std::rethrow_exception(callback_exception);
            }
            if (!finished_ireqs.empty()) {
                auto time_point = std::chrono::steady_clock::now();
                if (time_point > time_point_to_finish && latencies.size() > niter) {
                    break;
                }
                TimedIreq timedIreq = finished_ireqs.front();
                finished_ireqs.pop_front();
                lock.unlock();
                ov::InferRequest& ireq = timedIreq.ireq;
                if (timedIreq.has_start_time) {
                    latencies.push_back(std::chrono::duration_cast<Ms>(time_point - timedIreq.start).count());
                }
                ireq.set_callback(
                    [&ireq, time_point, &mutex, &finished_ireqs, &callback_exception, &cv](std::exception_ptr ex) {
                        // Keep callback small. This improves performance for fast (tens of thousands FPS) models
                        std::unique_lock<std::mutex> lock(mutex);
                        {
                            try {
                                if (ex) {
                                    std::rethrow_exception(ex);
                                }
                                finished_ireqs.push_back({ireq, time_point, true});
                            } catch (const std::exception&) {
                                if (!callback_exception) {
                                    callback_exception = std::current_exception();
                                }
                            }
                        }
                        cv.notify_one();
                    });
                ireq.start_async();
            }
        }
        auto end = std::chrono::steady_clock::now();
        double duration = std::chrono::duration_cast<Ms>(end - start).count();
        // Report results
        slog::info << "Count:      " << latencies.size() << " iterations" << slog::endl
                   << "Duration:   " << duration << " ms" << slog::endl
                   << "Latency:" << slog::endl;
        size_t percent = 50;
        LatencyMetrics{latencies, "", percent}.write_to_slog();
        slog::info << "Throughput: " << double_to_string(1000 * latencies.size() / duration) << " FPS" << slog::endl;
    } catch (const std::exception& ex) {
        slog::err << ex.what() << slog::endl;
        return EXIT_FAILURE;
    }
    return EXIT_SUCCESS;
}

You can see the explicit description of each sample step at Integration Steps section of “Integrate OpenVINO™ Runtime with Your Application” guide.

Running

python throughput_benchmark.py <path_to_model> <device_name>(default: CPU)

throughput_benchmark <path_to_model> <device_name>(default: CPU)

To run the sample, you need to specify a model. You can get a model specific for your inference task from one of model repositories, such as TensorFlow Zoo, HuggingFace, or TensorFlow Hub.

Example

  1. Download a pre-trained model.

  2. You can convert it by using:

    import openvino as ov

ov_model = ov.convert_model('./models/googlenet-v1')
# or, when model is a Python model object
ov_model = ov.convert_model(googlenet-v1)

    ovc ./models/googlenet-v1

  3. Perform benchmarking, using the googlenet-v1 model on a CPU:

    python throughput_benchmark.py ./models/googlenet-v1.xml

    throughput_benchmark ./models/googlenet-v1.xml

Sample Output

The application outputs performance results.

[ INFO ] OpenVINO:
[ INFO ] Build ................................. <version>
[ INFO ] Count:          2817 iterations
[ INFO ] Duration:       10012.65 ms
[ INFO ] Latency:
[ INFO ]     Median:     13.80 ms
[ INFO ]     Average:    14.10 ms
[ INFO ]     Min:        8.35 ms
[ INFO ]     Max:        28.38 ms
[ INFO ] Throughput: 281.34 FPS

The application outputs performance results.

[ INFO ] OpenVINO:
[ INFO ] Build ................................. <version>
[ INFO ] Count:      1577 iterations
[ INFO ] Duration:   15024.2 ms
[ INFO ] Latency:
[ INFO ]        Median:     38.02 ms
[ INFO ]        Average:    38.08 ms
[ INFO ]        Min:        25.23 ms
[ INFO ]        Max:        49.16 ms
[ INFO ] Throughput: 104.96 FPS